ARCHIVES OF ACOUSTICS Copyright c 2018 by PAN – IPPT Vol. 43, No. 3, pp. 505–516 (2018) DOI: 10.24425/123922 A Study on of Music Features Derived from Audio Recordings Examples – a Quantitative Analysis Aleksandra DOROCHOWICZ(1), Bożena KOSTEK(2) (1) Multimedia Systems Department Faculty of Electronics, Telecommunications and Informatics Gdansk University of Technology Narutowicza 11/12, 80-233 Gdańsk, Poland; e-mail: [email protected] (2) Audio Acoustics Laboratory Faculty of Electronics, Telecommunications and Informatics Gdansk University of Technology Narutowicza 11/12, 80-233 Gdańsk, Poland; e-mail: [email protected] (received June 19, 2017; accepted March 21, 2018) The paper presents a comparative study of music features derived from audio recordings, i.e. the same music pieces but representing different music genres, excerpts performed by different musicians, and songs performed by a musician, whose style evolved over time. Firstly, the origin and the background of the division of music genres were shortly presented. Then, several objective parameters of an audio signal were recalled that have an easy interpretation in the context of perceptual relevance. Within the study parameter values were extracted from music excerpts, gathered and compared to determine to what extent they are similar within the songs of the same performer or samples representing the same piece. Keywords: music genres; audio parametrization; music features. 1. Introduction account a number of feature types at the same time. Music genres are also divided into smaller sub-groups There are many formal systems describing differ- (punk rock, new wave, post grunge) and they can be ences between music pieces, which allow for defining mixed (symphonic metal, pop rock). its assignment to a category or type of music. Many Throughout the history, many types of classifica- features influence the final categorization of the song, tions were created. They were based on music styles, some of them can be described or measured as is forms and genres. Over time, the topologies become done within the Automatic music genre classification more and more complex and divided. Moreover, the (AMGC) area (Bergstra et al., 2006; Bhalke, 2017; pieces could belong to more than one category, so it is Burred, Lerch, 2014; Hoffmann, Kostek, 2015; necessary to identify their definitions and significance. Kalliris et al., 2016; Kotsakis et al., 2012; Kostek, These include musical form, style, genre and descrip- 2005; Ntalampiras, 2013; Schedl et al., 2014; Silla tive features such as meter, tempo, structure or origin. et al., 2007; Sturm, 2014; Tzanetakis et al., 2002; The definition of the music form is based on the vocal Rosner et al., 2014; Rosner, Kostek, 2018). One and/or music instruments used, texture type or num- of the ways the songs can be described is by using ber of passages it consists of (one-passaged, cyclic). their origin (J-rock, Brit pop), time when they are Moreover, the Small Music Encyclopedia describes mu- composed or performed, performance (Pluta et al., sic forms as the typical for a specific group of music 2017), mood of music (Barthet et al., 2017; Plewa, piece schemes, designated by the analysis based on the Kostek, 2015), instruments used (symphonic, acous- specific pieces. tic, rock), music techniques (riff, rap) or function that At the beginning of 17th century, the concept of music plays (film score, religious or orchestral music) the style in the music theory changed its meaning (Benward, 2003). Frequently, to be able to conduct (Pascall, 2001; Seidel, Leisinger, 1998). Division a proper analysis of a piece, one may need to take into based on antique tradition: religious, chamber and 506 Archives of Acoustics – Volume 43, Number 3, 2018 scenic (theatrical), within subtypes were proposed by et al., 2006; Tzanetakis et al., 2002; Holzapfel, Marc Scacchi in 1649 (Palisca, 1998). Athanasius Stylianou, 2008; Kostek et al., 2011; Kostek, Kircher proposed to expand it into church, madrigal Kaczmarek, 2013; Ntalampiras, 2013), where the and theatrical style, noting that it is not the place user may choose a song belonging to the particular of performance that specifies the style, but the ef- music genre (Tzanetakis et al., 2002). fects triggered. His proposal also predicts styles dif- The aim of the study is to determine to what extent fusions (Helman, 2016). Although the definition of music feature values and characteristics of music ex- the music style has changed through the centuries, cerpts are similar for the performer or for a music piece. according to the Small Music Encyclopedia, the mu- Tracks gathered for the purpose of this study are de- sic style is the term describing the common features scribed by their parametric representation, i.e. time of the compositional technique typical for the specific and spectral parameters, such as: RMS (Root-Mean- piece, author, for the national music, historical period Square) energy, zero-crossing rate, spectral centroid, (Dziębowska, 1998). spectral skewness, spectral kurtosis, spectral flatness, A music genre identifies some pieces of music as entropy of spectrum, brightness and roll-off, all ex- belonging to a shared tradition or set of conventions. tracted by the Matlab MIRtoolbox1.6.2 (MIRtoolbox). Music genres are created based on many types of di- At the same time several descriptive features such as visions: age of the recipients (e.g. adult contemporary, music genre, time of origin, country are identified. Mu- teen pop, music for children), artists’ origin (e.g. Amer- sic excerpts used in the experiment represent various icana, Afro-Cuban, Brit pop, Italo disco, K-pop, etc.), music genres, assigned according to the artist’s ambi- time of the origin (e.g. baroque, classical, contem- tions, origin, time of origin, they include various pieces porary, etc.), artists’/music ambitions (classical mu- of the same artist, and the same songs performed by sic, ambient, country, etc.), ideology (e.g. rock, hip- a number of various artists. hop, rap, metal, blues, New Age, etc.), instrumenta- tion and treatment of musical instruments within the given genre (e.g. jazz, country, folk, electronic, blues, 2. Experiments rock, etc.). 2.1. Building a database The American music magazines Alternative Press (Alternative Press, 2016) and Rock Sound (Rocksound, In the study, 202 music excerpts, 0.5–1 minute long, 2016) at the beginning of the 2016 drew attention to were collected. They represent various music genres, the problem of lack of the unequivocal music genres i.e.: rock, pop-punk, new wave, glam metal, punk rock, definition and divisions in the research study entitled Brit pop, pop, soft rock, blues, musical, rock and roll, “What is punk?”. It was carried out by the Converse psychedelic rock, soul, art rock, heavy metal, emo, company in associate with the Polygraph’s analyst post grunge, cabaret, pop rock, different time of ori- (Daniels, 2016). Their bases were playlists tagged as gin (from 1943 to 2016), different countries and instru- punk at Spotify and YouTube’s playlists. Over half of ments used. Also, the research includes samples of one them (51%) contain such music bands as: Green Day, song performed by various artists and various songs Blink-182 (50%), The Offspring (44%), Sum 41 (39%), performed by one artist. Rise Against (38%), Fall Out Boy (38%), My Chemi- Moreover, in our previous study we have performed cal Romance (35%), Bad Religion (30%), Nofx (28%), listening tests to eventually correlate the subjective All Time Low (28%) and A Day to Remember (27%). session outcome with the artificial intelligence algo- All the bands belong to genres closely related to punk rithms results (Dorochowicz et al., 2017). The task (pop-punk, emo, post-hardcore, metalcore), although of the test participant was to assign a music excerpt to they are not strictly punk. This research shows that it the specific music genre. The same goal was expected is not obvious to which category the performer or the to be achieved by the machine learning approach. Even song belongs, which makes defining it in some cases though the listening test results were not fully homo- almost impossible. geneous due to the individual characteristics of mu- It should also be noted that within the Music In- sic as well as the ambiguity of some tracks, causing formation Retrieval area there exists a considerable that people attributed a given track to different mu- disagreement among the researchers over music genre sic genres, still the results were more than encourag- classification, as assigned by a human or by a ma- ing (Dorochowicz et al., 2017). We have obtained chine learning algorithm, i.e. whether such an assign- a good agreement between subjective and “objective” ment is practical (Silla et al., 2017; Tekman, Hor- approaches, the latter one based on machine learning. tacsu, 2002). At the same time there are a lot of ex- This helps to determine which music excerpts can be amples of research studies, as well as commercial appli- considered as uniquely attributed to a particular genre, cations (e.g. music social networking, music catalogu- checked both by subjective tests and machine learn- ing tools for applications, etc.) related to automatic ing approach. Unambiguous cases obtained in that way genre recognition and classification (e.g. Bergsta were parametrized and further analysed. A. Dorochowicz, B. Kostek – A Study on of Music Features Derived from Audio Recordings Examples. 507 2.2. Parametrization • Spectral centroid – returns the first moment (mean), which is the geometric center (centroid) As already mentioned the analysis performed is (Fig. 2). It is interpreted as a measure of Bright- based on a quantitative comparison of parametric rep- ness (center of gravity; see Fig. 3): resentation of music excerpts. To parametrize music Z excerpts several features have been chosen, as defined µ = xf(x) dx: (2) by the Matlab MIRtoolbox 1.6.2 (MIRtoolbox).